Abstract: In order to solve the problem that a single algorithm feature descriptor is difficult to express both rigid and non-rigid 3D models, this paper propose a universal feature extraction method for 3D models. Firstly,a local area weighted densification sampling method based on 3D point cloud model is proposed. Secondly,aim at the influence of non-rigid hinge structure transformation, a time-scale sequence heat kernel encoding method is proposed by using the equidistant and isometric invariance of heat kernel signature. Finally, an edge-projection graph convolutional neural network which is applied to 3D model classification tasks is designed. This network learns the feature fusion on the spatial shape of the encoded point cloud and the time-scale sequence heat kernel. Experiments on the rigid 3D model dataset ModelNet40 and non-rigid 3D model dataset SHREC15 show that compared with the single rigid or non-rigid 3D model feature extraction methods, the proposed method can extract universal feature descriptors with significant discrimination, and the classification accuracy rates at 92.63% and 97.71%, respectively.